Convert averager rpc_aggregate_part to P2P

[justheuristic]

[this PR should be merged into libp2p branch]

Right now, the single heaviest communication pattern in hivemind is averaging model parameters with peers. This is currently done via AllreduceRunner in this RPC:
https://github.com/learning-at-home/hivemind/blob/master/hivemind/client/averaging/allreduce.py#L204

... called via this request method:
https://github.com/learning-at-home/hivemind/blob/master/hivemind/client/averaging/allreduce.py#L137

The main quest is to overhaul this method to work via P2P

• create a test where two AllreduceRunners are manually created with two connected hivemind.P2P instances
• call rpc_aggregate_part in allreduce patern, similarly to this test
• test that it still works with large tensors (>10k values)

Since gRPC is not optimized for large messages, we slice these tensors into parts using split_for_streaming/combine_from_streaming. That said, P2P transport may not need this partitioning.

[borzunov]

Solved in #323.